The Sensory Neuron as a Transformer: Permutation-Invariant Neural Networks for Reinforcement Learning